tech ethics

AI Companies Trained on Artists' Work. Now Everyone Is Arguing About Who Owns What.

What happened

The AI copyright question has escalated from theoretical to litigated to settled and back to unsettled in the space of 18 months. Anthropic reached a $1.5 billion settlement with authors in September 2025. In March 2026, Chicken Soup for the Soul filed suit against eight AI companies simultaneously, framing all training data use as a single act of mass infringement. Also in March, the US Supreme Court declined to hear a case about whether AI-generated content can itself be copyrighted, leaving lower courts split. On April 16, multiple analysts and outlets noted that the EU AI Act's training data disclosure requirements, now in their active phase, are creating compliance obligations that US law still does not require. The legal and regulatory landscape is fragmented: some companies settle, some litigate, and the underlying question of whether training on copyrighted data constitutes fair use remains unresolved in US courts.

AI companies trained on stolen work, some of them paid to make it go away, and the law still doesn't say they had to. The Anthropic settlement was not an admission; it was an insurance policy against a ruling that hasn't come yet.

The Hidden Bet

The fair use question will eventually be resolved by US courts, providing clarity

The Supreme Court's decision not to take the AI-generated copyright case signals it may continue to avoid AI copyright questions until Congress acts. Congress has shown no inclination to act. The result may not be eventual clarity but permanent legal ambiguity that functions as a de facto permission structure for the largest companies that can afford litigation and settlements while smaller competitors cannot.

Settlements like Anthropic's represent adequate compensation for creators

$1.5 billion sounds large but spread across the universe of authors whose work trained Claude, it represents cents per work per author. More importantly, it covers past use. It establishes no licensing mechanism for ongoing training. The next model generation will train on new data, and the question of compensation starts over.

The EU AI Act's training data requirements will create a global standard

The Act requires disclosure of what was trained on, not consent or payment. A company can comply with EU law by saying 'we trained on the Common Crawl dataset' without licensing a single piece of content. The transparency requirement is real; the compensation requirement is not.

The Real Disagreement

The core tension is between two things that both seem true: first, that AI models would not exist without the creative work of millions of people who were never asked permission and were never paid; second, that requiring consent and payment for training data at scale would make training models prohibitively expensive or impossible, in a way that might freeze AI development at its current state. Both are accurate statements. You cannot have both copyright protection and unconstrained AI development on the current timeline. Something has to give. The question is whether it is artists or AI companies, and who decides.

What No One Is Saying

The settlement wave, Anthropic, Udio, others, is being treated as evidence that the system is working. It is evidence of the opposite. Companies that know they are legally exposed are buying peace before a ruling that might cost more. The settlements insulate the current generation of models from liability while establishing no precedent. The next generation of models, trained on synthetic data or on licensed content from a small number of large content owners, will be even more concentrated in fewer hands and will represent an even smaller fraction of human creative output. The settlement economy is not resolving the copyright question; it is funding the next phase of the problem.

Who Pays

Individual artists, illustrators, and writers

Already happened; ongoing as each new model trains

Their work was ingested without consent. Settlements go to organizational plaintiffs and their lawyers. Individual creators receive nothing from Anthropic's $1.5 billion deal unless they were named plaintiffs.

Smaller AI companies and startups

Now and ongoing

The legal risk from training data is now large enough that only well-capitalized companies can absorb settlement costs. This creates a structural advantage for incumbents and a barrier to entry for challengers who lack $1.5 billion in settlement reserves.

Scenarios

Perpetual Ambiguity

US courts split further. No Supreme Court ruling. Congress debates but passes nothing. Companies settle selectively and continue training. The status quo, where large AI companies bear manageable legal risk and artists bear the costs, persists for the foreseeable future.

Signal SCOTUS denies certiorari on the second AI copyright case brought before it in 2026

Congressional Licensing Framework

Congress passes an AI training data licensing compulsory license scheme, modeled on music's Section 115. AI companies pay into a royalty pool. A collective licensing body distributes payments. Artists receive small but real amounts.

Signal The Senate Judiciary Committee holds hearings specifically on a training data licensing bill with bipartisan co-sponsors

Landmark Ruling Against Fair Use

A major circuit court rules that training on copyrighted material without license is not fair use. AI companies face retroactive liability for existing models. Settlement negotiations become existential for smaller firms. Large companies use their legal war chests to stay in the market while competitors cannot.

Signal The Ninth Circuit or Second Circuit issues a ruling in one of the pending cases that specifically rejects the fair use defense for training data

What Would Change This

If the EU AI Act's training data transparency requirements reveal that major models were trained on specific, identifiable works in volumes large enough to make the 'transformative use' argument implausible, that would shift the legal ground significantly. Or if a US circuit court issues a ruling on the merits rather than on standing or discovery, the ambiguity ends. Until one of those happens, the status quo favors AI companies.

power

OpenAI Wants Immunity. Anthropic Wants Accountability. One AI Bill Will Decide Which Vision Wins.

ethics

OpenAI Is Backing a Bill That Would Let It Cause Mass Death Without Full Liability. This Week It Got Sued for Causing a Stalking.

decision

Canada Wants to Label AI Content. The Problem Is Everything Is AI Content Now.